CN110728238A - Personnel re-detection method of fusion type neural network - Google Patents
Personnel re-detection method of fusion type neural network Download PDFInfo
- Publication number
- CN110728238A CN110728238A CN201910970957.1A CN201910970957A CN110728238A CN 110728238 A CN110728238 A CN 110728238A CN 201910970957 A CN201910970957 A CN 201910970957A CN 110728238 A CN110728238 A CN 110728238A
- Authority
- CN
- China
- Prior art keywords
- neural network
- image
- network
- fusion
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 29
- 230000004927 fusion Effects 0.000 title claims abstract description 22
- 238000001514 detection method Methods 0.000 title claims description 15
- 238000000605 extraction Methods 0.000 claims abstract description 28
- 238000000034 method Methods 0.000 claims abstract description 20
- 238000013145 classification model Methods 0.000 claims abstract description 9
- 238000013102 re-test Methods 0.000 claims abstract description 6
- 238000013527 convolutional neural network Methods 0.000 claims description 15
- 230000004913 activation Effects 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 6
- 238000012163 sequencing technique Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 230000011218 segmentation Effects 0.000 claims description 2
- 238000012549 training Methods 0.000 abstract description 18
- 230000000694 effects Effects 0.000 abstract description 6
- 238000011524 similarity measure Methods 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 11
- 238000005516 engineering process Methods 0.000 description 7
- 230000006872 improvement Effects 0.000 description 7
- 238000011161 development Methods 0.000 description 4
- 230000018109 developmental process Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 241000271897 Viperidae Species 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000005055 memory storage Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000036544 posture Effects 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004904 shortening Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a personnel retest method of a fusion type neural network, which aims at the problems of small data set and insufficient training samples and proposes to enlarge the data set by utilizing an improved generation type countermeasure network (GAN); then, the problem of feature extraction is effectively solved by using a scheme of combining a color model HSV and a scale-invariant local ternary pattern with neural network feature fusion; furthermore, in the classification model, the similarity measure is calculated using cross entropy loss. In addition, the neural network selects a residual error network with high convergence rate and good extraction effect, and a Dropout layer is added before the final convolution layer, aiming at preventing overfitting. The experimental result shows that when the number of samples is small, the characteristic is effectively extracted through the characteristic fusion scheme, and the method has a certain application value.
Description
Technical Field
The invention relates to the technical field of personnel retest, in particular to a personnel retest method of a fusion type neural network.
Background
With the increasing attention on cross-border tracking technology, people re-detection technology is also a big research hotspot. In practice, the challenge of human re-detection techniques is further increased by the influence of many external factors. In recent years, public safety is more and more emphasized, and camera monitoring is visible everywhere on a road. In security and law enforcement related applications, the ability to identify pedestrians in surveillance cameras is urgently needed. In the traditional method, the operation of identifying the pedestrian is performed by related operators, and the operation generally has manual design characteristics of edges, gradients and the like, which is called manual screening; however, in the world of the internet, the data size is getting larger and larger, and the manual screening method cannot meet the needs of people, so that it is expected to match pedestrians automatically and accurately in different cameras, which is generally called as people re-detection or pedestrian re-identification. In early studies, due to immaturity of related technologies such as image processing, pattern recognition and the like, the development of human re-detection technology has not been advanced; in recent years, the appearance of high-definition cameras and the development of image processing technology enable people to rapidly advance the re-detection technology. However, in the road of development progress, people re-detection technology faces great challenges. Shooting differences exist among different camera devices; meanwhile, pedestrians have diverse changes in movement, and the appearance of pedestrians is easily influenced by clothes, body types, postures, real object shielding, illumination backgrounds and the like.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to provide a personnel retest method of a fusion type neural network, which firstly aims at the problems of small data set and insufficient training samples, and proposes to enlarge the data set by using an improved generation type countermeasure network (GAN); then, the problem of feature extraction is effectively solved by using a scheme of combining a color model HSV and a scale-invariant local ternary pattern with neural network feature fusion; furthermore, in the classification model, the similarity measure is calculated using cross entropy loss. In addition, the neural network selects a residual error network with high convergence rate and good extraction effect, and a Dropout layer is added before the final convolution layer, aiming at preventing overfitting. The experimental result shows that when the number of samples is small, the characteristic is effectively extracted through the characteristic fusion scheme, and the method has a certain application value.
The invention provides a personnel re-detection method of a fusion type neural network, which comprises three parts, wherein the first part relates to a generative confrontation network, and a data set is expanded by utilizing the generative confrontation network GAN; the second part is a convolutional neural network, a generated image generated by using a GAN network is processed by using basic operations of convolution, pooling and function activation, an input image is extracted by combining additional manual features, and then a fusion layer is externally connected to fuse the convolutional neural network and the manually extracted features to obtain an image description feature with more complete representation; and the third part is to apply the minimized cross entropy to the classification model and perform precision sequencing on the human re-detection.
The generation type confrontation network (GAN) belongs to the unsupervised learning network, is rapidly applied to various computer vision fields, and has great success in deep learning and machine learning. The GAN includes a generative model G and a discriminative model D. The generator G learns how to generate an image from an input image, and the discriminator D learns whether the generated image is a real image or a "false image" generated by the generator G, both models learning in a mutual game, also called resist learning, so that the generated image gradually approximates the original image. To further improve the stability of GAN and the quality of generated pictures, two aspects can be considered: firstly, how to search a better model to train the GAN network; the second is how to improve the loss function in GAN. This results in improved network Deep Conditional General Adaptive Networks (DCGAN) and Least Square Generic Adaptive Networks (LSGAN), which use DCGAN in the present invention.
The further improvement lies in that: g and D in the GAN are two convolutional neural networks, the generator G is used for up-sampling by means of deconvolution, the discriminator D is used for replacing pooling by convolution added with stride, the generator G is provided with a Relu layer, and the discriminator D is provided with LeakyRelu. The first distinctive module is an improvement in detail compared to GAN, which replaces G and D in classical GAN with two Convolutional Neural Networks (CNN). However, instead of a direct substitution, upsampling is performed by deconvolution in generator G, posing is replaced by stride-added convolution in discriminator D, Relu layer is used in generator G, and leakyreu is used in discriminator D. First starting with a random vector of 100 dimensions and then scaling it up to 4 x 1024 with a linear function. To expand this tensor, a 4 step-2 micro-step convolution function is used, whose kernel size is 5 × 5. Each micro-stride convolution function is followed by a modified linear unit and a batch unit. In addition, there is an optional deconvolution layer with stride 1, whose kernel size is 5 × 5, and a Tanh activation function is added to fine-tune the result. Finally, a sample size of 64 × 64 × 3 can be generated. The resulting image may then be resized to 256 x 3 using a bilinear sampling method. The pictures are generated after DCGAN is trained on a face data set and a landscape data set, so that the picture set with very high quality can be generated under the condition of high reiteration times by using DCGAN, and the effect of expanding the data set can be achieved.
The training framework of DCGAN is used to train the images of the original data set, adjust all images to 128 x 48 in size before training, and flip them randomly to obtain more conformal data. The number of training iterations was set to 50, and 26000 images of 256 × 256 size were generated.
The feature extraction module is divided into two parts: traditional manual extraction of features and convolutional neural network extraction of features. The input image set is extracted using the conventional color descriptors HSV and the Scale Invariant Local Ternary Pattern (SILTP). SILTP is a well-known Local Binary Pattern (LBP) improvement operator, which has good invariance under monotonic gray scale transformation, but is not robust to image noise. SILTP improves LBP by introducing scale-invariant local comparison tolerances, achieving invariance to intensity scale variations and robustness to image noise.
The further improvement lies in that: in the first part, the image is preprocessed, the image is equivalently segmented, the image is divided into six strips in the horizontal direction, an image pyramid is constructed for each horizontal strip, and a single histogram is calculated, so that more complete multi-scale information is obtained. The resulting histogram achieves some invariance to viewpoint changes while capturing local regional features of the person. The specific operation is as follows: down-sampling of the original 128 x 48 image, with two 2 x 2 slipsAnd performing local average merging operation on the active cell window to describe local details of the person image, wherein the step length of the active cell window is 5 pixels, repeating the process, and finally obtaining two SILTP histograms and a combined HSV histogram of 8 multiplied by bin according to the calculated local maximum occurrence frequency, wherein each histogram box represents the occurrence probability of one mode in the active cell window. To account for viewpoint variations, all sub-windows at the same horizontal position are examined and the local occurrence of each mode in these sub-windows (i.e., the same histogram frame) is maximized. The resulting histogram achieves some invariance to viewpoint changes while capturing local regional features of the person. Using the obtained two single histograms as1-norm is standardized, and the two are connected into a whole to obtain a fusion Feature, which is called as a joint descriptor FFD (fusion Feature description) with the size of 960. However, spatial information within the bands may also be lost, thereby lacking the necessary characterization, affecting the resolution.
The further improvement lies in that: and generating a picture, and extracting features by using a residual error network ResNet.
Most body parts can be found in the image, but there are some severe distortions and misalignments, so the last fully-connected layer in the network uses higher-level convolutional layers, which consist of a stack of filters of predefined size, which are convolved with the input of the layer. Parameter sharing used by the convolutional layer is more efficient (requiring less computation and memory storage) than dense matrix multiplication. Parameter sharing may also make convolutional layers equivalent to linear translation (i.e., any shift in the input will result in a similar shift in the output). In addition, the pooling layer also adopts corresponding nonlinear expression, which is helpful for the normalization of feature extraction.
The DCGAN generated image data set uses the neural network ResNet18 to extract features, and deeper networks are not selected, so that the defects of high training requirement, long time and the like caused by gradient explosion prevention and increase of the number of layers are overcome. In addition, the residual error network can well solve the problems of degradation and the like through a residual error function, network optimization is achieved, and the convergence is quicker.
All training images in the present invention are adjusted to 256 × 256 size before neural network training, then randomly flipped horizontally and cropped to 224 × 224 size. When the neural network is trained, the learning rate is given to be 0.05, then the learning rate is reduced to be 0.001 after 40 times of iterative training, and the iterative training is carried out on a training set for 40 times respectively. The training strategy applies a small batch of random Gradient descent (sgd) (stochastic Gradient present) to modify the parameters to achieve faster back propagation and faster convergence. And a Dropout layer is added before the last convolutional layer to prevent over-fitting. And (4) drawing a training and checking loss curve, and showing that the training effect is basically optimal after 30-40 iterations.
The further improvement lies in that: the operator of manual feature extraction is combined with the CNN extraction operator, and through a fusion layer, two feature extraction channels are complemented to obtain a feature of 2048 tensor to represent image input as follows: x ═ HS, RN _ Features]The output is calculated by the following formula:
where h (-) represents the activation function.
The further improvement lies in that: a downsampling layer and a ReLU layer are adopted, the downsampling rate is set to be 0.5, the neural network adopts a feedback propagation algorithm, and the iteration formula is as follows:
the aim is to effectively extract complete features on each image, and not to extract features for comparison by using a neural network. Therefore, instead of using the verification model, a classification model is chosen. Efficient characterization results in lower losses. Here, a softmax loss function is selected for application in the model. For a single input vector x and a single output node last layer, the penalty can be calculated by the following formula:
the last layer of the network is designed to minimize cross-entropy loss:
the invention has the beneficial effects that: firstly, expanding a VIPeR data set by adopting an improved GAN network to obtain 26000 256 × 256 image sets, which can well reduce overfitting conditions for feature extraction; thirdly, combining manual extraction and neural network extraction, performing feature extraction on the input data set and the generated data set, and obtaining more complete feature expression through a fusion layer; and finally, sorting the similarity of the designated image and the images in the test set by calculating cross entropy loss through a classification model and adopting Rank-k evaluation indexes. Compared with the result of the traditional algorithm, the performance of the obtained result is improved to a certain extent. A learning method that operates by multiple layers of linearity and nonlinearity simultaneously learned in an end-to-end manner using a deep neural network. In order to accurately extract semantic features with good robustness, parameters of the layers are learned through multiple iterations. To expand the data set, a large number of high quality clear pictures are generated using a generative confrontation network GAN. Therefore, under the condition of a large number of levels of feature extraction, the initial method is effectively improved, and the performance is improved to a certain extent.
Drawings
FIG. 1 is a diagram of the Re-ID framework of the present invention.
Fig. 2 is a diagram of a DCGAN network architecture of the present invention.
Fig. 3 is a diagram of a residual network setup unit of the present invention.
FIG. 4 is a residual network training accuracy and check loss curve of the present invention.
Detailed Description
For the purpose of enhancing understanding of the present invention, the present invention will be further described in detail with reference to the following examples, which are provided for illustration only and are not to be construed as limiting the scope of the present invention. As shown in fig. 1-4, the present embodiment provides a method for re-detecting persons by using a fusion type neural network, which is divided into three parts, wherein the first part relates to a generative confrontation network, and a data set is expanded by using the generative confrontation network GAN; the second part is a convolutional neural network, a generated image generated by using a GAN network is processed by using basic operations of convolution, pooling and function activation, an input image is extracted by combining additional manual features, and then a fusion layer is externally connected to fuse the convolutional neural network and the manually extracted features to obtain an image description feature with more complete representation; and the third part is to apply the minimized cross entropy to the classification model and perform precision sequencing on the human re-detection.
G and D in the GAN are two convolutional neural networks, the generator G is used for up-sampling by means of deconvolution, the discriminator D is used for replacing pooling by convolution added with stride, the generator G is provided with a Relu layer, and the discriminator D is provided with LeakyRelu.
In the first part, the image is preprocessed, equivalent segmentation is carried out on the image, the image is divided into six strips in the horizontal direction, an image pyramid is constructed for each horizontal strip, a single histogram is calculated, the obtained histogram realizes certain invariance of viewpoint change, and meanwhile, local region characteristics of people are captured.
And generating a picture, and extracting features by using a residual error network ResNet.
The operator of manual feature extraction is combined with the CNN extraction operator, and through a fusion layer, two feature extraction channels are complemented to obtain a feature of 2048 tensor to represent image input as follows: x ═ HS, RN _ Features]The output is calculated by the following formula:
where h (-) represents the activation function.
A downsampling layer and a ReLU layer are adopted, the downsampling rate is set to be 0.5, the neural network adopts a feedback propagation algorithm, and the iteration formula is as follows:
the aim is to effectively extract complete features on each image, and not to extract features for comparison by using a neural network. Therefore, instead of using the verification model, a classification model is chosen. Efficient feature representation will result in lowerIs lost. Here, a softmax loss function is selected for application in the model. For a single input vector x and a single output node last layer, the penalty can be calculated by the following formula:the last layer of the network is designed to minimize cross-entropy loss:
compared with the traditional personnel re-detection algorithm, the method is correspondingly optimized when the network framework is constructed (1) when the image sets are few, useful information description is lacked in the feature extraction module, so that the phenomena of overfitting and the like are caused, the DCGAN network is adopted to generate a large number of pictures, the input data set is expanded, and the situation can be effectively eliminated; (2) the traditional manual feature extraction and neural network combination is adopted in a feature extraction module, the color information and the texture information are combined, more complete feature representation can be extracted, and a ResNet network is adopted, and extra convolutional layers, down-sampling and other operations are added, so that a better training effect can be achieved; (3) dropout is introduced into the classification model through the cross entropy model, so that the overfitting phenomenon can be reduced. Therefore, the training precision on the VIPeR data set is higher than that of the traditional algorithm, and the specific conditions are as follows in table 1:
table 1 individual person search evaluation table
According to the table, the proposed method combining the manual extraction and the neural network extraction features has better detection effect than the traditional feature learning and metric learning. The accuracy of conventional single feature extraction cannot meet the requirements. Further development can be continued on shortening the training running time in the future, and the obtained fusion features can be reprocessed.
Claims (6)
1. A personnel re-detection method of a fusion type neural network is characterized in that: the method is divided into three parts, wherein the first part relates to a generative confrontation network, and a data set is expanded by using the generative confrontation network GAN; the second part is a convolutional neural network, a generated image generated by using a GAN network is processed by using basic operations of convolution, pooling and function activation, an input image is extracted by combining additional manual features, and then a fusion layer is externally connected to fuse the convolutional neural network and the manually extracted features to obtain an image description feature with more complete representation; and the third part is to apply the minimized cross entropy to the classification model and perform precision sequencing on the human re-detection.
2. The method for retesting persons in a fused neural network as set forth in claim 1, wherein: g and D in the GAN are two convolutional neural networks, the generator G is used for up-sampling by means of deconvolution, the discriminator D is used for replacing pooling by convolution added with stride, the generator G is provided with a Relu layer, and the discriminator D is provided with LeakyRelu.
3. The method for retesting persons in a fused neural network as set forth in claim 1, wherein: in the first part, the image is preprocessed, equivalent segmentation is carried out on the image, the image is divided into six strips in the horizontal direction, an image pyramid is constructed for each horizontal strip, a single histogram is calculated, the obtained histogram realizes certain invariance of viewpoint change, and meanwhile, local region characteristics of people are captured.
4. The human retest method of the fused neural network of claim 3, wherein: and generating a picture, and extracting features by using a residual error network ResNet.
5. The method for retesting persons in a fused neural network as set forth in claim 1, wherein: the operator of manual feature extraction is combined with the CNN extraction operator, and two feature extraction channels are complemented through a fusion layer to obtain a feature of 2048 tensor to representThe image input is: x ═ HS, RN _ Features]The output is calculated by the following formula:
where h (-) represents the activation function.
6. The human retest method of the fused neural network of claim 5, wherein: a downsampling layer and a ReLU layer are adopted, the downsampling rate is set to be 0.5, the neural network adopts a feedback propagation algorithm, and the iteration formula is as follows:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910970957.1A CN110728238A (en) | 2019-10-12 | 2019-10-12 | Personnel re-detection method of fusion type neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910970957.1A CN110728238A (en) | 2019-10-12 | 2019-10-12 | Personnel re-detection method of fusion type neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110728238A true CN110728238A (en) | 2020-01-24 |
Family
ID=69220076
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910970957.1A Pending CN110728238A (en) | 2019-10-12 | 2019-10-12 | Personnel re-detection method of fusion type neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110728238A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112037223A (en) * | 2020-11-06 | 2020-12-04 | 中科创达软件股份有限公司 | Image defect detection method and device and electronic equipment |
| CN112230210A (en) * | 2020-09-09 | 2021-01-15 | 南昌航空大学 | HRRP radar target identification method based on improved LSGAN and CNN |
| CN113240097A (en) * | 2021-06-08 | 2021-08-10 | 西安邮电大学 | Data expansion and classification method and system |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108960141A (en) * | 2018-07-04 | 2018-12-07 | 国家新闻出版广电总局广播科学研究院 | Pedestrian's recognition methods again based on enhanced depth convolutional neural networks |
| US20180374233A1 (en) * | 2017-06-27 | 2018-12-27 | Qualcomm Incorporated | Using object re-identification in video surveillance |
| WO2019007004A1 (en) * | 2017-07-04 | 2019-01-10 | 北京大学深圳研究生院 | Image feature extraction method for person re-identification |
| WO2019010950A1 (en) * | 2017-07-13 | 2019-01-17 | 北京大学深圳研究生院 | Depth discrimination network model method for pedestrian re-recognition in image or video |
| CN109446898A (en) * | 2018-09-20 | 2019-03-08 | 暨南大学 | A kind of recognition methods again of the pedestrian based on transfer learning and Fusion Features |
| CN109829918A (en) * | 2019-01-02 | 2019-05-31 | 安徽工程大学 | A kind of liver image dividing method based on dense feature pyramid network |
| CN109902590A (en) * | 2019-01-30 | 2019-06-18 | 西安理工大学 | Pedestrian's recognition methods again of depth multiple view characteristic distance study |
| CN109961051A (en) * | 2019-03-28 | 2019-07-02 | 湖北工业大学 | A kind of pedestrian's recognition methods again extracted based on cluster and blocking characteristic |
| CN110046599A (en) * | 2019-04-23 | 2019-07-23 | 东北大学 | Intelligent control method based on depth integration neural network pedestrian weight identification technology |
| WO2019144575A1 (en) * | 2018-01-24 | 2019-08-01 | 中山大学 | Fast pedestrian detection method and device |
| CN110084108A (en) * | 2019-03-19 | 2019-08-02 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Pedestrian re-identification system and method based on GAN neural network |
-
2019
- 2019-10-12 CN CN201910970957.1A patent/CN110728238A/en active Pending
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180374233A1 (en) * | 2017-06-27 | 2018-12-27 | Qualcomm Incorporated | Using object re-identification in video surveillance |
| WO2019007004A1 (en) * | 2017-07-04 | 2019-01-10 | 北京大学深圳研究生院 | Image feature extraction method for person re-identification |
| WO2019010950A1 (en) * | 2017-07-13 | 2019-01-17 | 北京大学深圳研究生院 | Depth discrimination network model method for pedestrian re-recognition in image or video |
| WO2019144575A1 (en) * | 2018-01-24 | 2019-08-01 | 中山大学 | Fast pedestrian detection method and device |
| CN108960141A (en) * | 2018-07-04 | 2018-12-07 | 国家新闻出版广电总局广播科学研究院 | Pedestrian's recognition methods again based on enhanced depth convolutional neural networks |
| CN109446898A (en) * | 2018-09-20 | 2019-03-08 | 暨南大学 | A kind of recognition methods again of the pedestrian based on transfer learning and Fusion Features |
| CN109829918A (en) * | 2019-01-02 | 2019-05-31 | 安徽工程大学 | A kind of liver image dividing method based on dense feature pyramid network |
| CN109902590A (en) * | 2019-01-30 | 2019-06-18 | 西安理工大学 | Pedestrian's recognition methods again of depth multiple view characteristic distance study |
| CN110084108A (en) * | 2019-03-19 | 2019-08-02 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Pedestrian re-identification system and method based on GAN neural network |
| CN109961051A (en) * | 2019-03-28 | 2019-07-02 | 湖北工业大学 | A kind of pedestrian's recognition methods again extracted based on cluster and blocking characteristic |
| CN110046599A (en) * | 2019-04-23 | 2019-07-23 | 东北大学 | Intelligent control method based on depth integration neural network pedestrian weight identification technology |
Non-Patent Citations (5)
| Title |
|---|
| JING-TAO WANG 等: "Pedestrian recognition in multi-camera networks based on deep transfer learning and feature visualization" * |
| 洪洋 等: "深度卷积对抗生成网络综述" * |
| 熊炜 等: "基于CNN的改进行人重识别技术" * |
| 王俊茜: "基于多任务联合监督学习的行人再识别研究" * |
| 韦忠亮 等: "基于深度学习和距离度量的行人再识别研究" * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112230210A (en) * | 2020-09-09 | 2021-01-15 | 南昌航空大学 | HRRP radar target identification method based on improved LSGAN and CNN |
| CN112037223A (en) * | 2020-11-06 | 2020-12-04 | 中科创达软件股份有限公司 | Image defect detection method and device and electronic equipment |
| CN112037223B (en) * | 2020-11-06 | 2021-03-26 | 中科创达软件股份有限公司 | Image defect detection method and device and electronic equipment |
| CN113240097A (en) * | 2021-06-08 | 2021-08-10 | 西安邮电大学 | Data expansion and classification method and system |
| CN113240097B (en) * | 2021-06-08 | 2024-04-26 | 西安邮电大学 | Method and system for expanding and classifying data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bui et al. | Using grayscale images for object recognition with convolutional-recursive neural network | |
| Abbas et al. | Region-based object detection and classification using faster R-CNN | |
| CN108875624B (en) | Face detection method based on multi-scale cascade dense connection neural network | |
| Sabzmeydani et al. | Detecting pedestrians by learning shapelet features | |
| US10198657B2 (en) | All-weather thermal-image pedestrian detection method | |
| CN110929593B (en) | Real-time significance pedestrian detection method based on detail discrimination | |
| Le et al. | Deeply Supervised 3D Recurrent FCN for Salient Object Detection in Videos. | |
| Anwar et al. | Depth Estimation and Blur Removal from a Single Out-of-focus Image. | |
| CN106529494A (en) | Human face recognition method based on multi-camera model | |
| CN113688894B (en) | Fine granularity image classification method integrating multiple granularity features | |
| CN113011357A (en) | Depth fake face video positioning method based on space-time fusion | |
| Zhang et al. | License plate localization in unconstrained scenes using a two-stage CNN-RNN | |
| CN112036260B (en) | Expression recognition method and system for multi-scale sub-block aggregation in natural environment | |
| CN110728238A (en) | Personnel re-detection method of fusion type neural network | |
| Saqib et al. | Person head detection in multiple scales using deep convolutional neural networks | |
| Cao et al. | Learning spatial-temporal representation for smoke vehicle detection | |
| CN116563410A (en) | Electrical equipment electric spark image generation method based on two-stage generation countermeasure network | |
| Sharma et al. | Deepfakes Classification of Faces Using Convolutional Neural Networks. | |
| CN113706404B (en) | Depression angle face image correction method and system based on self-attention mechanism | |
| Wang et al. | Robust object representation by boosting-like deep learning architecture | |
| Huang et al. | Multi-Teacher Single-Student Visual Transformer with Multi-Level Attention for Face Spoofing Detection. | |
| KR20010050988A (en) | Scale and Rotation Invariant Intelligent Face Detection | |
| Chacon-Murguia et al. | Moving object detection in video sequences based on a two-frame temporal information CNN | |
| KR20180092453A (en) | Face recognition method Using convolutional neural network and stereo image | |
| CN111242114A (en) | Character recognition method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200124 |